基于用户反馈的多策略翻译在线融合方法研究

项目名称： 基于用户反馈的多策略翻译在线融合方法研究

项目编号： No.61272384

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 杨沐昀

作者单位： 哈尔滨工业大学

项目金额： 80万元

中文摘要： 随着互联网时代对机器翻译的需求凸显，如何利用现有翻译技术满足广泛灵活的用户需求已成为机器翻译研究的巨大挑战。多策略机器翻译方法旨在融合多种翻译模型的优势，是解决这一问题的合理对策。与现有系统融合不同，本课题将多策略翻译视为一种满足多变用户需求的在线学习问题。课题首先以译文融合特征挖掘为基础，尝试解决译文融合时使用的特征与翻译建模过程同构的问题；进而分析翻译系统用户的行为，从中量化用户满意度而不单纯是译文质量作为融合目标，以解决现有自动翻译评价倾向于SMT结果的偏置问题；最终针对复杂多样的实际翻译需求，引入在线学习机制，探索实现基于在线排序学习的多策略翻译融合方法。课题目的是研究适用于机器翻译问题的在线机器学习方法，实现一个性能良好的基于在线学习的多策略机器翻译模型，为建立一种能够综合利用多种翻译模型和翻译知识以满足用户个性需求的多策略翻译机理进行有益的探索。

中文关键词： 翻译反馈；多策略；在线学习；；

英文摘要： With the pressing demands on translation service brought by the Internet, the MT technology is faced with the challenge to meet the complex and flexible end user needs. The hybrid MT is a promising solution to this issue owing to its pursuing to combine the merits of various existing MT methods. In contrast to the up-to-date system combination technology, this project treats the hybrid MT as an online learning task to combine proper MT results for different users' needs. Starting from an examination on current system combination techniques, this project first tries to resolve the duplicate feature space issue between the system combination and the translation modeling process via feature mining. Then, it explores to build a user-centered rather than translation-quality-centered measure via user feedbacks for translation optimization, so as to eliminate the bias for SMT among current automatic evaluation metrics. Finally, the project investigates into the construction of an online ranking model to enable an dynamic multi-strategy MT process to different real user needs. In addition, the attainments of this project would be informative on how to combine the multiple translation knowledges and translation models for a personalized MT result.

英文关键词： Translation Feedback；Hybrid Approach；Online Learning；；